TARS-EEG: a tropical algebraic representation for EEG seizure detection

This is the authors' abstract. We don't add key points for this paper.

Objectives To propose an interpretable tropical algebraic signal representation, named TARS-EEG, for EEG-based seizure detection by modelling short EEG segments through max-plus and min-plus waveform-dominance profiles. Methods Short EEG segments from the Epileptic Seizure Recognition dataset were mapped into multiscale tropical dilation, erosion, gradient, and residue profiles. A Tropical Seizure Index was introduced to quantify normalized local waveform-dominance energy. The extracted tropical features were evaluated using standard classifiers with recording-group-aware train-test analysis, recording-group-aware five-fold cross-validation, and ablation experiments. Results The proposed TARS-EEG representation achieved stable seizure/non-seizure classification performance across standard classifiers. Under recording-group-aware five-fold cross-validation, MLP achieved the highest mean accuracy of 97.70 %, while XGBoost achieved the highest ROC-AUC of 0.9926. Ablation analysis showed that the complete multiscale tropical representation performed better than individual neighbourhood-scale representations. Conclusions TARS-EEG provides a compact and interpretable algebraic representation for EEG seizure detection. The results suggest that multiscale max-plus and min-plus waveform descriptors can capture seizure-related local waveform dominance effectively.